DNV is developing an automatic corrosion detection solution which facilitates decision support for inspection activities. It employs deep learning-based image recognition algorithms and a low-cost depth camera to detect corrosion and measure its extent in various scenarios
Corrosion prevention is expensive due to the harsh marine en[ds_preview]vironment. A recent study estimates that »the total cost of marine and offshore corrosion worldwide is between USD 50–80 billion«. Manufacturers often use high-performance coatings to prevent corrosion, but these will eventually show signs of corrosion. The marine industry therefore complies with IACS Recommendation 87 (IACS (2015)), which divides the coating condition into three grades: Good, Fair and Poor, to assess the coating condition.
Currently, the coating condition is assessed manually by surveyors – not only DNV’s but also those of other classification societies. This raises two types of challenges: i) surveyors have to be physically present onboard to carry out their inspection, which is likely to lead to hazardous situations; ii) the assessment of the coating condition relies highly on the surveyor’s expertise and ability, thus involving the innate weakness of the human vision system in unfavourable environments and making the assessment prone to inaccuracy and inconsistency.
Classification societies and other marine service providers have long striven to develop (semi-)automated inspection tools in an effort to boost efficiency and reduce costs. This work has recently accelerated as part of the ongoing digital transformation. However, current state-of-the-art methods assess the condition of the coating on metal surfaces by using a RGB camera as a sensor. This method lacks effectiveness because the size of a detected corrosion area solely depends on the number of corresponding pixels, without considering the object distance and orientation. To build a coating condition assessment tool that is robust against changes in appearance, a promising solution is to involve 3D information combined with RGB images. The ability to capture 3D information is currently provided by a so-called depth sensor. In the current prototype stage, our proposed corrosion detection solution uses an Intel RealSense D435i camera. This RealSense device provides hardware-level synchronized RGB and depth images. The next wave of innovation will bring depth sensors as standard to mobile handsets as well as drones. With this technology, our two-tiered corrosion condition assessment method works as follows: firstly, a deep learning-based image segmentation model identifies areas of corrosion (as well as coating breakdown); secondly, the detected corrosion area can be re-projected to 3D world coordinates to determine its physical size. This ability enables a more reliable assessment of the corrosion condition and improves the state of the art.
Successful results
The proposed automatic corrosion condition assessment solution consists of two independent software components. The first component is the Corrosion Detection Framework (CDF), which identifies pixels representing corrosion from an image. After this processing, each pixel will be classified as either »corrosion« or »non-corrosion« (in the case of an external structure, pixels can be additionally classified as »non-region-of-interest« due to the effect of Non-RoI removal, as described below). Then the second component, the Area Calculation Module (ACM) uses the identified set of pixels as input to calculate the area. The two modules will be explained in further detail in the following sections.
The Corrosion Detection Framework (CDF) is composed of three steps: i) Scene classification step – given an input image, a fine-tuned classification network with MobileNetV2 architecture will first categorize it as one of the two defined scenarios, i.e., internal structure and external structure. ii) Non-RoI Removal step – in the case of an external structure, an additional step named Non-RoI Removal is conducted to remove irrelevant image regions. iii) Corrosion segmentation step – the CDF uses U-Net to segment corrosion in the input image. More specifically, two independently trained U-Net models are provided to deal with images of ship internal and external structures respectively.
Area Calculation Module (ACM). According to IACS Recommendation 87, the coating condition is assessed based upon the estimated percentage of areas with coating failure and rusted surfaces. A straightforward way to constitute corrosion area is to count the number of pixels classified as corrosion in an image. In our work, ACM directly calculates the physical sizes with the help of depth information.
The method proposed in this paper lends efficient and effective coating condition assessment support to surveyors, shipyards, and other maritime stakeholders. The output of our method is the ratio of the corroded area(s) to the whole tank surface, which helps surveyors make rapid decisions. Once fully implemented and coupled with an inspection drone, this novel approach helps to reduce the efforts required and can make inspection operations safer.
Authors: Qian Wei, Yanzhi Chen, DNV Artificial Intelligence Research Centre, Shanghai
